Process data analytics and monitoring based on causality analysis techniques

2014-08-06A typical industrial process or plant operates with hundreds of control loops and those primary loops should operate at desired levels for safety and efficiency. There exist challenges on how to monitor and maintain such a large‐scale complex system, thus control performance monitoring (CPM) has drawn tremendous research interest and progress steadily over the last two decades. After detection of poor control performance, an even more challenging and meaningful task is to diagnose and locate the root cause of whole plant degradation therefore further targeted maintenance could be performed to improve the situation. In this dissertation, new data‐driven approaches based on causality methods are proposed to diagnose the root cause of plant‐wide disturbances. ❧ Ocillations as a common reason for poor performance in closed‐loop controlled processes which, once generated, can propagate along process flows and feedback paths of the whole plant. A new data‐driven time series method for diagnosing the sources and propagation paths of plant‐wide oscillations is presented. Firstly a latent variable based feature variable selection scheme is proposed to select candidate variables which share common oscillations. The time‐domain Granger causality and spectral Granger causality have been applied successfully to reveal the causal relationship in terms of time series prediction. ❧ The proposed method is then extended to extract more complete causal structure in a large‐scale complex industrial plant. A novel multilevel Granger causality framework for root cause diagnosis is proposed, which consider both the group‐wise and within group causal structures. Dynamic time warping‐based K‐means clustering method is applied to group time‐series variables more accurately. Group Granger causal test is performed to find out the causal contribution with regard to specific disturbance feature. Dominant and major causal relationship could also be extracted in group point of view. Moreover, a partial least squares modified Granger causal test is developed to overcome multicollinearity issue and make the generated graphic causal network much sparse and easier to interpret. ❧ For nonstationary series in chemical process, traditional latent variable or other statistical modeling methods are inadequate because the statistical properties are time variant. Nonstationary tests are implemented to identify the nonstationary variables, then cointegration test are utilized to capture the common trends and dynamic structure. Monitoring index and control limit are derived for monitoring and fault detection purpose